Dynamic frequency modulation linearization for small coherent spectrometers via transfer learning

With the wide application of high-speed swept-frequency laser technology in precision measurement and spectral analysis, the non-linearity of laser frequency modulation and dynamic environmental disturbances restrict system performance. The traditional static look-up table fails to model dynamic effects such as current jumps and temperature drifts, whereas conventional dynamic compensation lacks real-time performance and cross-device generalization. To address this, this paper proposes a dynamic frequency correction method integrating transfer learning and a dual-branch network: it constructs static/dynamic feature subnets to extract prior current-frequency mapping and real-time disturbance response, designs residual calibration and temperature adaptive modules for multi-source feature fusion, and introduces an adversarial fine-tuning mechanism based on the fast gradient sign method to enhance robustness against distribution shifts. Experimental results verify its advantages in delay, precision, and power consumption, providing an effective solution for high-linearity swept-frequency light sources in small coherent spectrometers.